Separate Lanczos_PRO, add sketch of memory profiling

Author: lorenzotomada

.gitignore

@@ -174,16 +174,22 @@ cython_debug/
 *.dat
 *.data
 
+# other data files and logs
+data/*
+logs/*
+
 # temporary folders used for profiling
 tmp_data/*
+*tmp*
 
 # for documentation
 docs/_*
 
-#vscode files
+# vscode files
 .vscode/*
 
-#skbuild folder
+# skbuild folder
 skbuild/*
-data/*
+
+# temporary files for vim users
 *.swp

experiments/config.yaml

@@ -0,0 +1,2 @@
+max_iter: 1000
+tol: 0.0001

experiments/config_accuracy.yaml

@@ -0,0 +1,2 @@
+dim: 20000
+density: 0.2

experiments/config_profiling.yaml

@@ -1,78 +1,117 @@
-import cupyx.scipy.sparse as cpsp
-import cupy as cp
-from pyclassify import (
-    eigenvalues_np,
-    eigenvalues_sp,
-    # eigenvalues_cp,
-    power_method,
-    power_method_numba,
-    # power_method_cp,
-    QR,
-    # QR_cp,
-)
-from pyclassify.utils import (
-    make_symmetric,
-    read_config,
-    profile_with_cprofile,
-    # profile_with_cupy_profiler,
-)
+from pyclassify import power_method, power_method_numba
+from pyclassify.utils import make_symmetric, read_config
+from pyclassify.profiling_MPI import mpi_profiled, get_memory_usage_mb, profile_serial
 import numpy as np
 import scipy.sparse as sp
-import scipy
 import random
 import argparse
+import os
+import time
+import pandas as pd
+from mpi4py import MPI
 
 
+# Seed for reproducibility
 seed = 8422
 random.seed(seed)
-# cp.random.seed(seed)
 np.random.seed(seed)
 
+# Some MPI info
+comm = MPI.COMM_WORLD
+rank = comm.Get_rank()
+n_procs = comm.Get_size()
 
+
+# Here we parse the arguments. We provide a default value, but the user is free to chose another config file.
 parser = argparse.ArgumentParser()
 parser.add_argument("--config", type=str, required=False, help="config file:")
-
 args = parser.parse_args()
-filename = (
-    args.config if args.config else "./experiments/config"
-)  # automatic choice if no argument is passed
+config_file = args.config if args.config else "./experiments/config_profiling"
+
+
+# Now read the config only if the rank is 0. We will broadcast the info to all other ranks.
+if rank == 0:
+    kwargs = read_config(config_file)
+    dim = kwargs["dim"]
+    density = kwargs["density"]
+else:
+    dim = None
+    density = None
+
+dim = comm.bcast(dim, root=0)
+density = comm.bcast(density, root=0)
+
+
+# Generate the data
+if rank == 0:
+    A = sp.random(dim, dim, density=density, format="csr")
+    A = make_symmetric(A)
+else:
+    A = None
+
+A = comm.bcast(A, root=0)
+
+
+# Now we start profiling. Notice that the only function that requires MPI is the one that is not profiled within a 'if rank==0' statement.
+
+# @mpi_profiled
+# def profiled_divide_et_impera(A):
+#    from pyclassify import divide_et_impera  # avoid circular import
+#    return divide_et_impera(A)
+
+results = {}
+
+
+if rank == 0:
+    _ = profile_serial(power_method_numba, A.toarray())
+    results["power_method"] = profile_serial(power_method, A)
+    results["power_method_numba"] = profile_serial(power_method_numba, A.toarray())
+    # results["QR"] = profile_serial(QR, A)
 
+# mpi_result_QR = profiled_QR(A)
+# mpi_result_divide = profiled_divide_et_impera(A)
 
-kwargs = read_config(filename)
-dim = kwargs["dim"]
-density = kwargs["density"]
-tol = kwargs["tol"]
-max_iter = kwargs["max_iter"]
+# if rank == 0:
+#    results["QR"] = mpi_result_QR
+#    results["divide_et_impera"] = mpi_result_divide
 
 
-eigenvals = np.arange(1, dim + 1)
-A = np.diag(eigenvals)
-U = scipy.stats.ortho_group.rvs(dim)
-A = U @ A @ U.T
-A = make_symmetric(A)
-A = sp.csr_matrix(A)
-# A_cp = cpsp.csr_matrix(A)
+# Now we just save to CSV.
+if rank == 0:
+    os.makedirs("logs", exist_ok=True)
+    mem_csv = "logs/memory.csv"
+    time_csv = "logs/time.csv"
 
+    mem_row = {
+        "matrix_size": dim,
+        "density": density,
+        "num_procs": n_procs,
+        **{key: results[key]["memory_total"] for key in results},
+    }
 
-log_file = "./logs/timings.csv"
-iteration_factor = 300
+    time_row = {
+        "matrix_size": dim,
+        "density": density,
+        "num_procs": n_procs,
+        **{key: results[key]["time"] for key in results},
+    }
 
+    def append_or_create_csv(path, row, columns):
+        """
+        This helper function just decides whether to append to an existing CSV or to create a new one.
+        """
+        if not os.path.exists(path):
+            pd.DataFrame([row]).to_csv(path, index=False, columns=columns)
+        else:
+            df = pd.read_csv(path)
+            df = pd.concat([df, pd.DataFrame([row])], ignore_index=True)
+            df.to_csv(path, index=False, columns=columns)
 
-profile_with_cprofile(
-    log_file, dim, "eigenvalues_np", eigenvalues_np, A.toarray(), symmetric=True
-)
-profile_with_cprofile(
-    log_file, dim, "eigenvalues_sp", eigenvalues_sp, A, symmetric=True
-)
-profile_with_cprofile(log_file, dim, "power_method", power_method, A)
-profile_with_cprofile(
-    log_file, dim, "power_method_numba", power_method_numba, A.toarray()
-)
-profile_with_cprofile(
-    log_file, dim, "QR", QR, A.toarray(), max_iter=iteration_factor * dim
-)
+    method_names = list(results.keys())
+    base_cols = ["matrix_size", "density", "num_procs"]
+    all_columns = base_cols + method_names
 
+    append_or_create_csv(mem_csv, mem_row, all_columns)
+    append_or_create_csv(time_csv, time_row, all_columns)
 
-# profile_with_cupy_profiler(log_file, dim, "eigenvalues_cp", eigenvalues_cp, A_cp)
-# profile_with_cupy_profiler(log_file, dim, "power_method_cp", power_method_cp, A_cp)
-# profile_with_cupy_profiler(log_file, dim, "QR_cp", QR_cp, A_cp, q0=cp.random.rand(dim), tol=1e-3, max_iter=iteration_factor * dim)
+    print("Done! The results have been saved to logs/memory.csv and logs/time.csv")

scripts/profiling.py

@@ -1,81 +0,0 @@
-# import cupyx.scipy.sparse as cpsp
-# import cupy as cp
-from pyclassify import (
-    eigenvalues_np,
-    eigenvalues_sp,
-    # eigenvalues_cp,
-    power_method,
-    power_method_numba,
-    # power_method_cp,
-    QR,
-    # QR_cp,
-)
-from pyclassify.utils import (
-    make_symmetric,
-    read_config,
-    profile_with_cprofile,
-    # profile_with_cupy_profiler,
-)
-import numpy as np
-import scipy.sparse as sp
-import scipy
-import random
-import argparse
-
-
-# cp.cuda.Device(0).use()
-# cp.get_default_memory_pool().free_all_blocks()
-
-
-seed = 8422
-random.seed(seed)
-# cp.random.seed(seed)
-np.random.seed(seed)
-
-
-parser = argparse.ArgumentParser()
-parser.add_argument("--config", type=str, required=False, help="config file:")
-
-args = parser.parse_args()
-filename = (
-    args.config if args.config else "./experiments/config"
-)  # automatic choice if no argument is passed
-
-
-kwargs = read_config(filename)
-dim = kwargs["dim"]
-density = kwargs["density"]
-tol = kwargs["tol"]
-max_iter = kwargs["max_iter"]
-
-
-eigenvals = np.arange(1, dim + 1)
-A = np.diag(eigenvals)
-U = scipy.stats.ortho_group.rvs(dim)
-A = U @ A @ U.T
-A = make_symmetric(A)
-A = sp.csr_matrix(A)
-# A_cp = cpsp.csr_matrix(A)
-
-
-log_file = "./logs/timings.csv"
-iteration_factor = 300
-
-
-profile_with_cprofile(
-    log_file, dim, "eigenvalues_np", eigenvalues_np, A.toarray(), symmetric=True
-)
-profile_with_cprofile(
-    log_file, dim, "eigenvalues_sp", eigenvalues_sp, A, symmetric=True
-)
-profile_with_cprofile(log_file, dim, "power_method", power_method, A)
-profile_with_cprofile(
-    log_file, dim, "power_method_numba", power_method_numba, A.toarray()
-)
-profile_with_cprofile(
-    log_file, dim, "QR", QR, A.toarray(), max_iter=iteration_factor * dim
-)
-
-# profile_with_cupy_profiler(log_file, dim, "eigenvalues_cp", eigenvalues_cp, A_cp)
-# profile_with_cupy_profiler(log_file, dim, "power_method_cp", power_method_cp, A_cp)
-# profile_with_cupy_profiler(log_file, dim, "QR_cp", QR_cp, A_cp, q0=cp.random.rand(dim), tol=1e-3, max_iter=iteration_factor * dim)

scripts/run.py

@@ -1,3 +1 @@
 #!/bin/bash
-python -m kernprof -l -o logs/profile_eigenvalues.dat scripts/run.py --config=experiments/config
-python -m line_profiler -rmt "logs/profile_eigenvalues.dat" > logs/eigenvalues.txt

shell/profile.sh

@@ -1,4 +1,4 @@
+#!/bin/bash
 module load cmake/3.29.1
 module load intel/2021.2
-# Comment the following line if you are not interested in running using MPI
 module load openmpi3/3.1.4

shell/submit.sh

@@ -1,29 +1,21 @@
 __all__ = [
     "eigenvalues_np",
     "eigenvalues_sp",
-    # "eigenvalues_cp",
     "power_method",
     "power_method_numba",
-    # "power_method_cp",
+    "Lanczos_PRO",
     "EigenSolver",
-    # "Lanczos_PRO_cp",
-    # "QR_method_cp",
-    # "QR_cp",
 ]
 
 from .QR_cpp import QR_algorithm, Eigen_value_calculator
 
 from .eigenvalues import (
     eigenvalues_np,
     eigenvalues_sp,
-    # eigenvalues_cp,
     power_method,
     power_method_numba,
-    # power_method_cp,
+    Lanczos_PRO,
     EigenSolver,
-    # Lanczos_PRO_cp,
-    # QR_method_cp,
-    # QR_cp,
 )
 
-from .zero_finder import compute_Psi, secular_solver
+from .zero_finder import compute_Psi, secular_solver